Reproduce “Towards Audio Codec-based Speech Separation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jia Qi Yip, Shengkui Zhao, Dianwen Ng, Eng Siong Chng, Bin Ma
Codecformer, a new speech separation method, enables efficient, on-device processing by operating within the compressed embedding space of a pre-trained neural audio codec, resulting in a 52x reduction in inference computational cost compared to traditional models while maintaining competitive separation performance.
Key insights
- Performing speech separation within the compressed embedding space of a pre-trained NAC is feasible and leads to substantial computational efficiency gains without a drastic drop in performance.
- The significant temporal compression provided by NACs allows for simplified SS model architectures, eliminating the need for complex chunking mechanisms common in waveform-based models.
- The Residual Vector Quantization (RVQ) module within the chosen NAC introduces notable signal degradation that negatively impacts speech separation performance, suggesting areas for codec improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.